20 * or visit www.oracle.com if you need additional information or have any |
20 * or visit www.oracle.com if you need additional information or have any |
21 * questions. |
21 * questions. |
22 */ |
22 */ |
23 |
23 |
24 #include "precompiled.hpp" |
24 #include "precompiled.hpp" |
25 #include "opto/castnode.hpp" |
25 #include "classfile/javaClasses.hpp" |
|
26 #include "gc/z/c2/zBarrierSetC2.hpp" |
|
27 #include "gc/z/zBarrierSet.hpp" |
|
28 #include "gc/z/zBarrierSetAssembler.hpp" |
|
29 #include "gc/z/zBarrierSetRuntime.hpp" |
|
30 #include "opto/block.hpp" |
26 #include "opto/compile.hpp" |
31 #include "opto/compile.hpp" |
27 #include "opto/escape.hpp" |
|
28 #include "opto/graphKit.hpp" |
32 #include "opto/graphKit.hpp" |
29 #include "opto/loopnode.hpp" |
|
30 #include "opto/machnode.hpp" |
33 #include "opto/machnode.hpp" |
31 #include "opto/macro.hpp" |
|
32 #include "opto/memnode.hpp" |
34 #include "opto/memnode.hpp" |
33 #include "opto/movenode.hpp" |
|
34 #include "opto/node.hpp" |
35 #include "opto/node.hpp" |
35 #include "opto/phase.hpp" |
36 #include "opto/regalloc.hpp" |
36 #include "opto/phaseX.hpp" |
|
37 #include "opto/rootnode.hpp" |
37 #include "opto/rootnode.hpp" |
38 #include "opto/type.hpp" |
|
39 #include "utilities/copy.hpp" |
|
40 #include "utilities/growableArray.hpp" |
38 #include "utilities/growableArray.hpp" |
41 #include "utilities/macros.hpp" |
39 #include "utilities/macros.hpp" |
42 #include "gc/z/zBarrierSet.hpp" |
40 |
43 #include "gc/z/c2/zBarrierSetC2.hpp" |
41 class ZBarrierSetC2State : public ResourceObj { |
44 #include "gc/z/zThreadLocalData.hpp" |
42 private: |
45 #include "gc/z/zBarrierSetRuntime.hpp" |
43 GrowableArray<ZLoadBarrierStubC2*>* _stubs; |
46 |
44 Node_Array _live; |
47 ZBarrierSetC2State::ZBarrierSetC2State(Arena* comp_arena) : |
45 |
48 _load_barrier_nodes(new (comp_arena) GrowableArray<LoadBarrierNode*>(comp_arena, 8, 0, NULL)) {} |
46 public: |
49 |
47 ZBarrierSetC2State(Arena* arena) : |
50 int ZBarrierSetC2State::load_barrier_count() const { |
48 _stubs(new (arena) GrowableArray<ZLoadBarrierStubC2*>(arena, 8, 0, NULL)), |
51 return _load_barrier_nodes->length(); |
49 _live(arena) {} |
52 } |
50 |
53 |
51 GrowableArray<ZLoadBarrierStubC2*>* stubs() { |
54 void ZBarrierSetC2State::add_load_barrier_node(LoadBarrierNode * n) { |
52 return _stubs; |
55 assert(!_load_barrier_nodes->contains(n), " duplicate entry in expand list"); |
53 } |
56 _load_barrier_nodes->append(n); |
54 |
57 } |
55 RegMask* live(const Node* node) { |
58 |
56 if (!node->is_Mach()) { |
59 void ZBarrierSetC2State::remove_load_barrier_node(LoadBarrierNode * n) { |
57 // Don't need liveness for non-MachNodes |
60 // this function may be called twice for a node so check |
58 return NULL; |
61 // that the node is in the array before attempting to remove it |
59 } |
62 if (_load_barrier_nodes->contains(n)) { |
60 |
63 _load_barrier_nodes->remove(n); |
61 const MachNode* const mach = node->as_Mach(); |
64 } |
62 if (mach->barrier_data() != ZLoadBarrierStrong && |
65 } |
63 mach->barrier_data() != ZLoadBarrierWeak) { |
66 |
64 // Don't need liveness data for nodes without barriers |
67 LoadBarrierNode* ZBarrierSetC2State::load_barrier_node(int idx) const { |
65 return NULL; |
68 return _load_barrier_nodes->at(idx); |
66 } |
|
67 |
|
68 RegMask* live = (RegMask*)_live[node->_idx]; |
|
69 if (live == NULL) { |
|
70 live = new (Compile::current()->comp_arena()->Amalloc_D(sizeof(RegMask))) RegMask(); |
|
71 _live.map(node->_idx, (Node*)live); |
|
72 } |
|
73 |
|
74 return live; |
|
75 } |
|
76 }; |
|
77 |
|
78 static ZBarrierSetC2State* barrier_set_state() { |
|
79 return reinterpret_cast<ZBarrierSetC2State*>(Compile::current()->barrier_set_state()); |
|
80 } |
|
81 |
|
82 ZLoadBarrierStubC2* ZLoadBarrierStubC2::create(const MachNode* node, Address ref_addr, Register ref, Register tmp, bool weak) { |
|
83 ZLoadBarrierStubC2* const stub = new (Compile::current()->comp_arena()) ZLoadBarrierStubC2(node, ref_addr, ref, tmp, weak); |
|
84 if (!Compile::current()->in_scratch_emit_size()) { |
|
85 barrier_set_state()->stubs()->append(stub); |
|
86 } |
|
87 |
|
88 return stub; |
|
89 } |
|
90 |
|
91 ZLoadBarrierStubC2::ZLoadBarrierStubC2(const MachNode* node, Address ref_addr, Register ref, Register tmp, bool weak) : |
|
92 _node(node), |
|
93 _ref_addr(ref_addr), |
|
94 _ref(ref), |
|
95 _tmp(tmp), |
|
96 _weak(weak), |
|
97 _entry(), |
|
98 _continuation() { |
|
99 assert_different_registers(ref, ref_addr.base()); |
|
100 assert_different_registers(ref, ref_addr.index()); |
|
101 } |
|
102 |
|
103 Address ZLoadBarrierStubC2::ref_addr() const { |
|
104 return _ref_addr; |
|
105 } |
|
106 |
|
107 Register ZLoadBarrierStubC2::ref() const { |
|
108 return _ref; |
|
109 } |
|
110 |
|
111 Register ZLoadBarrierStubC2::tmp() const { |
|
112 return _tmp; |
|
113 } |
|
114 |
|
115 address ZLoadBarrierStubC2::slow_path() const { |
|
116 const DecoratorSet decorators = _weak ? ON_WEAK_OOP_REF : ON_STRONG_OOP_REF; |
|
117 return ZBarrierSetRuntime::load_barrier_on_oop_field_preloaded_addr(decorators); |
|
118 } |
|
119 |
|
120 RegMask& ZLoadBarrierStubC2::live() const { |
|
121 return *barrier_set_state()->live(_node); |
|
122 } |
|
123 |
|
124 Label* ZLoadBarrierStubC2::entry() { |
|
125 // The _entry will never be bound when in_scratch_emit_size() is true. |
|
126 // However, we still need to return a label that is not bound now, but |
|
127 // will eventually be bound. Any lable will do, as it will only act as |
|
128 // a placeholder, so we return the _continuation label. |
|
129 return Compile::current()->in_scratch_emit_size() ? &_continuation : &_entry; |
|
130 } |
|
131 |
|
132 Label* ZLoadBarrierStubC2::continuation() { |
|
133 return &_continuation; |
69 } |
134 } |
70 |
135 |
71 void* ZBarrierSetC2::create_barrier_state(Arena* comp_arena) const { |
136 void* ZBarrierSetC2::create_barrier_state(Arena* comp_arena) const { |
72 return new(comp_arena) ZBarrierSetC2State(comp_arena); |
137 return new (comp_arena) ZBarrierSetC2State(comp_arena); |
73 } |
138 } |
74 |
139 |
75 ZBarrierSetC2State* ZBarrierSetC2::state() const { |
140 void ZBarrierSetC2::late_barrier_analysis() const { |
76 return reinterpret_cast<ZBarrierSetC2State*>(Compile::current()->barrier_set_state()); |
141 analyze_dominating_barriers(); |
77 } |
142 compute_liveness_at_stubs(); |
78 |
143 } |
79 bool ZBarrierSetC2::is_gc_barrier_node(Node* node) const { |
144 |
80 // 1. This step follows potential oop projections of a load barrier before expansion |
145 void ZBarrierSetC2::emit_stubs(CodeBuffer& cb) const { |
81 if (node->is_Proj()) { |
146 MacroAssembler masm(&cb); |
82 node = node->in(0); |
147 GrowableArray<ZLoadBarrierStubC2*>* const stubs = barrier_set_state()->stubs(); |
83 } |
148 |
84 |
149 for (int i = 0; i < stubs->length(); i++) { |
85 // 2. This step checks for unexpanded load barriers |
150 // Make sure there is enough space in the code buffer |
86 if (node->is_LoadBarrier()) { |
151 if (cb.insts()->maybe_expand_to_ensure_remaining(Compile::MAX_inst_size) && cb.blob() == NULL) { |
87 return true; |
152 ciEnv::current()->record_failure("CodeCache is full"); |
88 } |
153 return; |
89 |
154 } |
90 // 3. This step checks for the phi corresponding to an optimized load barrier expansion |
155 |
91 if (node->is_Phi()) { |
156 ZBarrierSet::assembler()->generate_c2_load_barrier_stub(&masm, stubs->at(i)); |
92 PhiNode* phi = node->as_Phi(); |
157 } |
93 Node* n = phi->in(1); |
158 |
94 if (n != NULL && n->is_LoadBarrierSlowReg()) { |
159 masm.flush(); |
95 return true; |
160 } |
96 } |
161 |
97 } |
162 int ZBarrierSetC2::estimate_stub_size() const { |
98 |
163 Compile* const C = Compile::current(); |
99 return false; |
164 BufferBlob* const blob = C->scratch_buffer_blob(); |
100 } |
165 GrowableArray<ZLoadBarrierStubC2*>* const stubs = barrier_set_state()->stubs(); |
101 |
166 int size = 0; |
102 void ZBarrierSetC2::register_potential_barrier_node(Node* node) const { |
167 |
103 if (node->is_LoadBarrier()) { |
168 for (int i = 0; i < stubs->length(); i++) { |
104 state()->add_load_barrier_node(node->as_LoadBarrier()); |
169 CodeBuffer cb(blob->content_begin(), (address)C->scratch_locs_memory() - blob->content_begin()); |
105 } |
170 MacroAssembler masm(&cb); |
106 } |
171 ZBarrierSet::assembler()->generate_c2_load_barrier_stub(&masm, stubs->at(i)); |
107 |
172 size += cb.insts_size(); |
108 void ZBarrierSetC2::unregister_potential_barrier_node(Node* node) const { |
173 } |
109 if (node->is_LoadBarrier()) { |
174 |
110 state()->remove_load_barrier_node(node->as_LoadBarrier()); |
175 return size; |
111 } |
|
112 } |
|
113 |
|
114 void ZBarrierSetC2::eliminate_useless_gc_barriers(Unique_Node_List &useful, Compile* C) const { |
|
115 // Remove useless LoadBarrier nodes |
|
116 ZBarrierSetC2State* s = state(); |
|
117 for (int i = s->load_barrier_count()-1; i >= 0; i--) { |
|
118 LoadBarrierNode* n = s->load_barrier_node(i); |
|
119 if (!useful.member(n)) { |
|
120 unregister_potential_barrier_node(n); |
|
121 } |
|
122 } |
|
123 } |
|
124 |
|
125 void ZBarrierSetC2::enqueue_useful_gc_barrier(PhaseIterGVN* igvn, Node* node) const { |
|
126 if (node->is_LoadBarrier() && !node->as_LoadBarrier()->has_true_uses()) { |
|
127 igvn->_worklist.push(node); |
|
128 } |
|
129 } |
|
130 |
|
131 const uint NoBarrier = 0; |
|
132 const uint RequireBarrier = 1; |
|
133 const uint WeakBarrier = 2; |
|
134 const uint ExpandedBarrier = 4; |
|
135 |
|
136 static bool load_require_barrier(LoadNode* load) { return (load->barrier_data() & RequireBarrier) == RequireBarrier; } |
|
137 static bool load_has_weak_barrier(LoadNode* load) { return (load->barrier_data() & WeakBarrier) == WeakBarrier; } |
|
138 static bool load_has_expanded_barrier(LoadNode* load) { return (load->barrier_data() & ExpandedBarrier) == ExpandedBarrier; } |
|
139 static void load_set_expanded_barrier(LoadNode* load) { return load->set_barrier_data(ExpandedBarrier); } |
|
140 |
|
141 static void load_set_barrier(LoadNode* load, bool weak) { |
|
142 if (weak) { |
|
143 load->set_barrier_data(RequireBarrier | WeakBarrier); |
|
144 } else { |
|
145 load->set_barrier_data(RequireBarrier); |
|
146 } |
|
147 } |
|
148 |
|
149 // == LoadBarrierNode == |
|
150 |
|
151 LoadBarrierNode::LoadBarrierNode(Compile* C, |
|
152 Node* c, |
|
153 Node* mem, |
|
154 Node* val, |
|
155 Node* adr, |
|
156 bool weak) : |
|
157 MultiNode(Number_of_Inputs), |
|
158 _weak(weak) { |
|
159 init_req(Control, c); |
|
160 init_req(Memory, mem); |
|
161 init_req(Oop, val); |
|
162 init_req(Address, adr); |
|
163 init_req(Similar, C->top()); |
|
164 |
|
165 init_class_id(Class_LoadBarrier); |
|
166 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); |
|
167 bs->register_potential_barrier_node(this); |
|
168 } |
|
169 |
|
170 uint LoadBarrierNode::size_of() const { |
|
171 return sizeof(*this); |
|
172 } |
|
173 |
|
174 bool LoadBarrierNode::cmp(const Node& n) const { |
|
175 ShouldNotReachHere(); |
|
176 return false; |
|
177 } |
|
178 |
|
179 const Type *LoadBarrierNode::bottom_type() const { |
|
180 const Type** floadbarrier = (const Type **)(Compile::current()->type_arena()->Amalloc_4((Number_of_Outputs)*sizeof(Type*))); |
|
181 Node* in_oop = in(Oop); |
|
182 floadbarrier[Control] = Type::CONTROL; |
|
183 floadbarrier[Memory] = Type::MEMORY; |
|
184 floadbarrier[Oop] = in_oop == NULL ? Type::TOP : in_oop->bottom_type(); |
|
185 return TypeTuple::make(Number_of_Outputs, floadbarrier); |
|
186 } |
|
187 |
|
188 const TypePtr* LoadBarrierNode::adr_type() const { |
|
189 ShouldNotReachHere(); |
|
190 return NULL; |
|
191 } |
|
192 |
|
193 const Type *LoadBarrierNode::Value(PhaseGVN *phase) const { |
|
194 const Type** floadbarrier = (const Type **)(phase->C->type_arena()->Amalloc_4((Number_of_Outputs)*sizeof(Type*))); |
|
195 const Type* val_t = phase->type(in(Oop)); |
|
196 floadbarrier[Control] = Type::CONTROL; |
|
197 floadbarrier[Memory] = Type::MEMORY; |
|
198 floadbarrier[Oop] = val_t; |
|
199 return TypeTuple::make(Number_of_Outputs, floadbarrier); |
|
200 } |
|
201 |
|
202 bool LoadBarrierNode::is_dominator(PhaseIdealLoop* phase, bool linear_only, Node *d, Node *n) { |
|
203 if (phase != NULL) { |
|
204 return phase->is_dominator(d, n); |
|
205 } |
|
206 |
|
207 for (int i = 0; i < 10 && n != NULL; i++) { |
|
208 n = IfNode::up_one_dom(n, linear_only); |
|
209 if (n == d) { |
|
210 return true; |
|
211 } |
|
212 } |
|
213 |
|
214 return false; |
|
215 } |
|
216 |
|
217 LoadBarrierNode* LoadBarrierNode::has_dominating_barrier(PhaseIdealLoop* phase, bool linear_only, bool look_for_similar) { |
|
218 if (is_weak()) { |
|
219 // Weak barriers can't be eliminated |
|
220 return NULL; |
|
221 } |
|
222 |
|
223 Node* val = in(LoadBarrierNode::Oop); |
|
224 if (in(Similar)->is_Proj() && in(Similar)->in(0)->is_LoadBarrier()) { |
|
225 LoadBarrierNode* lb = in(Similar)->in(0)->as_LoadBarrier(); |
|
226 assert(lb->in(Address) == in(Address), ""); |
|
227 // Load barrier on Similar edge dominates so if it now has the Oop field it can replace this barrier. |
|
228 if (lb->in(Oop) == in(Oop)) { |
|
229 return lb; |
|
230 } |
|
231 // Follow chain of load barrier through Similar edges |
|
232 while (!lb->in(Similar)->is_top()) { |
|
233 lb = lb->in(Similar)->in(0)->as_LoadBarrier(); |
|
234 assert(lb->in(Address) == in(Address), ""); |
|
235 } |
|
236 if (lb != in(Similar)->in(0)) { |
|
237 return lb; |
|
238 } |
|
239 } |
|
240 for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) { |
|
241 Node* u = val->fast_out(i); |
|
242 if (u != this && u->is_LoadBarrier() && u->in(Oop) == val && u->as_LoadBarrier()->has_true_uses()) { |
|
243 Node* this_ctrl = in(LoadBarrierNode::Control); |
|
244 Node* other_ctrl = u->in(LoadBarrierNode::Control); |
|
245 if (is_dominator(phase, linear_only, other_ctrl, this_ctrl)) { |
|
246 return u->as_LoadBarrier(); |
|
247 } |
|
248 } |
|
249 } |
|
250 |
|
251 if (can_be_eliminated()) { |
|
252 return NULL; |
|
253 } |
|
254 |
|
255 if (!look_for_similar) { |
|
256 return NULL; |
|
257 } |
|
258 |
|
259 Node* addr = in(LoadBarrierNode::Address); |
|
260 for (DUIterator_Fast imax, i = addr->fast_outs(imax); i < imax; i++) { |
|
261 Node* u = addr->fast_out(i); |
|
262 if (u != this && u->is_LoadBarrier() && u->as_LoadBarrier()->has_true_uses()) { |
|
263 Node* this_ctrl = in(LoadBarrierNode::Control); |
|
264 Node* other_ctrl = u->in(LoadBarrierNode::Control); |
|
265 if (is_dominator(phase, linear_only, other_ctrl, this_ctrl)) { |
|
266 ResourceMark rm; |
|
267 Unique_Node_List wq; |
|
268 wq.push(in(LoadBarrierNode::Control)); |
|
269 bool ok = true; |
|
270 bool dom_found = false; |
|
271 for (uint next = 0; next < wq.size(); ++next) { |
|
272 Node *n = wq.at(next); |
|
273 if (n->is_top()) { |
|
274 return NULL; |
|
275 } |
|
276 assert(n->is_CFG(), ""); |
|
277 if (n->is_SafePoint()) { |
|
278 ok = false; |
|
279 break; |
|
280 } |
|
281 if (n == u) { |
|
282 dom_found = true; |
|
283 continue; |
|
284 } |
|
285 if (n->is_Region()) { |
|
286 for (uint i = 1; i < n->req(); i++) { |
|
287 Node* m = n->in(i); |
|
288 if (m != NULL) { |
|
289 wq.push(m); |
|
290 } |
|
291 } |
|
292 } else { |
|
293 Node* m = n->in(0); |
|
294 if (m != NULL) { |
|
295 wq.push(m); |
|
296 } |
|
297 } |
|
298 } |
|
299 if (ok) { |
|
300 assert(dom_found, ""); |
|
301 return u->as_LoadBarrier(); |
|
302 } |
|
303 break; |
|
304 } |
|
305 } |
|
306 } |
|
307 |
|
308 return NULL; |
|
309 } |
|
310 |
|
311 void LoadBarrierNode::push_dominated_barriers(PhaseIterGVN* igvn) const { |
|
312 // Change to that barrier may affect a dominated barrier so re-push those |
|
313 assert(!is_weak(), "sanity"); |
|
314 Node* val = in(LoadBarrierNode::Oop); |
|
315 |
|
316 for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) { |
|
317 Node* u = val->fast_out(i); |
|
318 if (u != this && u->is_LoadBarrier() && u->in(Oop) == val) { |
|
319 Node* this_ctrl = in(Control); |
|
320 Node* other_ctrl = u->in(Control); |
|
321 if (is_dominator(NULL, false, this_ctrl, other_ctrl)) { |
|
322 igvn->_worklist.push(u); |
|
323 } |
|
324 } |
|
325 |
|
326 Node* addr = in(LoadBarrierNode::Address); |
|
327 for (DUIterator_Fast imax, i = addr->fast_outs(imax); i < imax; i++) { |
|
328 Node* u = addr->fast_out(i); |
|
329 if (u != this && u->is_LoadBarrier() && u->in(Similar)->is_top()) { |
|
330 Node* this_ctrl = in(Control); |
|
331 Node* other_ctrl = u->in(Control); |
|
332 if (is_dominator(NULL, false, this_ctrl, other_ctrl)) { |
|
333 igvn->_worklist.push(u); |
|
334 } |
|
335 } |
|
336 } |
|
337 } |
|
338 } |
|
339 |
|
340 Node *LoadBarrierNode::Identity(PhaseGVN *phase) { |
|
341 LoadBarrierNode* dominating_barrier = has_dominating_barrier(NULL, true, false); |
|
342 if (dominating_barrier != NULL) { |
|
343 assert(!is_weak(), "Weak barriers cant be eliminated"); |
|
344 assert(dominating_barrier->in(Oop) == in(Oop), ""); |
|
345 return dominating_barrier; |
|
346 } |
|
347 |
|
348 return this; |
|
349 } |
|
350 |
|
351 Node *LoadBarrierNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
|
352 if (remove_dead_region(phase, can_reshape)) { |
|
353 return this; |
|
354 } |
|
355 |
|
356 Node *val = in(Oop); |
|
357 Node *mem = in(Memory); |
|
358 Node *ctrl = in(Control); |
|
359 |
|
360 assert(val->Opcode() != Op_LoadN, ""); |
|
361 assert(val->Opcode() != Op_DecodeN, ""); |
|
362 |
|
363 if (mem->is_MergeMem()) { |
|
364 Node *new_mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw); |
|
365 set_req(Memory, new_mem); |
|
366 if (mem->outcnt() == 0 && can_reshape) { |
|
367 phase->is_IterGVN()->_worklist.push(mem); |
|
368 } |
|
369 return this; |
|
370 } |
|
371 |
|
372 LoadBarrierNode *dominating_barrier = NULL; |
|
373 if (!is_weak()) { |
|
374 dominating_barrier = has_dominating_barrier(NULL, !can_reshape, !phase->C->major_progress()); |
|
375 if (dominating_barrier != NULL && dominating_barrier->in(Oop) != in(Oop)) { |
|
376 assert(in(Address) == dominating_barrier->in(Address), ""); |
|
377 set_req(Similar, dominating_barrier->proj_out(Oop)); |
|
378 return this; |
|
379 } |
|
380 } |
|
381 |
|
382 bool eliminate = can_reshape && (dominating_barrier != NULL || !has_true_uses()); |
|
383 if (eliminate) { |
|
384 if (can_reshape) { |
|
385 PhaseIterGVN* igvn = phase->is_IterGVN(); |
|
386 Node* out_ctrl = proj_out_or_null(Control); |
|
387 Node* out_res = proj_out_or_null(Oop); |
|
388 |
|
389 if (out_ctrl != NULL) { |
|
390 igvn->replace_node(out_ctrl, ctrl); |
|
391 } |
|
392 |
|
393 // That transformation may cause the Similar edge on the load barrier to be invalid |
|
394 fix_similar_in_uses(igvn); |
|
395 if (out_res != NULL) { |
|
396 if (dominating_barrier != NULL) { |
|
397 assert(!is_weak(), "Sanity"); |
|
398 igvn->replace_node(out_res, dominating_barrier->proj_out(Oop)); |
|
399 } else { |
|
400 igvn->replace_node(out_res, val); |
|
401 } |
|
402 } |
|
403 } |
|
404 return new ConINode(TypeInt::ZERO); |
|
405 } |
|
406 |
|
407 // If the Similar edge is no longer a load barrier, clear it |
|
408 Node* similar = in(Similar); |
|
409 if (!similar->is_top() && !(similar->is_Proj() && similar->in(0)->is_LoadBarrier())) { |
|
410 set_req(Similar, phase->C->top()); |
|
411 return this; |
|
412 } |
|
413 |
|
414 if (can_reshape && !is_weak()) { |
|
415 // If this barrier is linked through the Similar edge by a |
|
416 // dominated barrier and both barriers have the same Oop field, |
|
417 // the dominated barrier can go away, so push it for reprocessing. |
|
418 // We also want to avoid a barrier to depend on another dominating |
|
419 // barrier through its Similar edge that itself depend on another |
|
420 // barrier through its Similar edge and rather have the first |
|
421 // depend on the third. |
|
422 PhaseIterGVN* igvn = phase->is_IterGVN(); |
|
423 Node* out_res = proj_out(Oop); |
|
424 for (DUIterator_Fast imax, i = out_res->fast_outs(imax); i < imax; i++) { |
|
425 Node* u = out_res->fast_out(i); |
|
426 if (u->is_LoadBarrier() && u->in(Similar) == out_res && |
|
427 (u->in(Oop) == val || !u->in(Similar)->is_top())) { |
|
428 assert(!u->as_LoadBarrier()->is_weak(), "Sanity"); |
|
429 igvn->_worklist.push(u); |
|
430 } |
|
431 } |
|
432 push_dominated_barriers(igvn); |
|
433 } |
|
434 |
|
435 return NULL; |
|
436 } |
|
437 |
|
438 uint LoadBarrierNode::match_edge(uint idx) const { |
|
439 ShouldNotReachHere(); |
|
440 return 0; |
|
441 } |
|
442 |
|
443 void LoadBarrierNode::fix_similar_in_uses(PhaseIterGVN* igvn) { |
|
444 Node* out_res = proj_out_or_null(Oop); |
|
445 if (out_res == NULL) { |
|
446 return; |
|
447 } |
|
448 |
|
449 for (DUIterator_Fast imax, i = out_res->fast_outs(imax); i < imax; i++) { |
|
450 Node* u = out_res->fast_out(i); |
|
451 if (u->is_LoadBarrier() && u->in(Similar) == out_res) { |
|
452 igvn->replace_input_of(u, Similar, igvn->C->top()); |
|
453 --i; |
|
454 --imax; |
|
455 } |
|
456 } |
|
457 } |
|
458 |
|
459 bool LoadBarrierNode::has_true_uses() const { |
|
460 Node* out_res = proj_out_or_null(Oop); |
|
461 if (out_res != NULL) { |
|
462 for (DUIterator_Fast imax, i = out_res->fast_outs(imax); i < imax; i++) { |
|
463 Node *u = out_res->fast_out(i); |
|
464 if (!u->is_LoadBarrier() || u->in(Similar) != out_res) { |
|
465 return true; |
|
466 } |
|
467 } |
|
468 } |
|
469 return false; |
|
470 } |
176 } |
471 |
177 |
472 static bool barrier_needed(C2Access& access) { |
178 static bool barrier_needed(C2Access& access) { |
473 return ZBarrierSet::barrier_needed(access.decorators(), access.type()); |
179 return ZBarrierSet::barrier_needed(access.decorators(), access.type()); |
474 } |
180 } |
475 |
181 |
476 Node* ZBarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const { |
182 Node* ZBarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const { |
477 Node* p = BarrierSetC2::load_at_resolved(access, val_type); |
183 Node* result = BarrierSetC2::load_at_resolved(access, val_type); |
478 if (!barrier_needed(access)) { |
184 if (barrier_needed(access) && access.raw_access()->is_Mem()) { |
479 return p; |
185 if ((access.decorators() & ON_WEAK_OOP_REF) != 0) { |
480 } |
186 access.raw_access()->as_Load()->set_barrier_data(ZLoadBarrierWeak); |
481 |
187 } else { |
482 bool weak = (access.decorators() & ON_WEAK_OOP_REF) != 0; |
188 access.raw_access()->as_Load()->set_barrier_data(ZLoadBarrierStrong); |
483 if (p->isa_Load()) { |
189 } |
484 load_set_barrier(p->as_Load(), weak); |
190 } |
485 } |
191 |
486 return p; |
192 return result; |
487 } |
193 } |
488 |
194 |
489 Node* ZBarrierSetC2::atomic_cmpxchg_val_at_resolved(C2AtomicParseAccess& access, Node* expected_val, |
195 Node* ZBarrierSetC2::atomic_cmpxchg_val_at_resolved(C2AtomicParseAccess& access, Node* expected_val, |
490 Node* new_val, const Type* val_type) const { |
196 Node* new_val, const Type* val_type) const { |
491 Node* result = BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, val_type); |
197 Node* result = BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, val_type); |
492 LoadStoreNode* lsn = result->as_LoadStore(); |
|
493 if (barrier_needed(access)) { |
198 if (barrier_needed(access)) { |
494 lsn->set_has_barrier(); |
199 access.raw_access()->as_LoadStore()->set_barrier_data(ZLoadBarrierStrong); |
495 } |
200 } |
496 return lsn; |
201 return result; |
497 } |
202 } |
498 |
203 |
499 Node* ZBarrierSetC2::atomic_cmpxchg_bool_at_resolved(C2AtomicParseAccess& access, Node* expected_val, |
204 Node* ZBarrierSetC2::atomic_cmpxchg_bool_at_resolved(C2AtomicParseAccess& access, Node* expected_val, |
500 Node* new_val, const Type* value_type) const { |
205 Node* new_val, const Type* value_type) const { |
501 Node* result = BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type); |
206 Node* result = BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type); |
502 LoadStoreNode* lsn = result->as_LoadStore(); |
|
503 if (barrier_needed(access)) { |
207 if (barrier_needed(access)) { |
504 lsn->set_has_barrier(); |
208 access.raw_access()->as_LoadStore()->set_barrier_data(ZLoadBarrierStrong); |
505 } |
209 } |
506 return lsn; |
210 return result; |
507 } |
211 } |
508 |
212 |
509 Node* ZBarrierSetC2::atomic_xchg_at_resolved(C2AtomicParseAccess& access, Node* new_val, const Type* val_type) const { |
213 Node* ZBarrierSetC2::atomic_xchg_at_resolved(C2AtomicParseAccess& access, Node* new_val, const Type* val_type) const { |
510 Node* result = BarrierSetC2::atomic_xchg_at_resolved(access, new_val, val_type); |
214 Node* result = BarrierSetC2::atomic_xchg_at_resolved(access, new_val, val_type); |
511 LoadStoreNode* lsn = result->as_LoadStore(); |
|
512 if (barrier_needed(access)) { |
215 if (barrier_needed(access)) { |
513 lsn->set_has_barrier(); |
216 access.raw_access()->as_LoadStore()->set_barrier_data(ZLoadBarrierStrong); |
514 } |
217 } |
515 return lsn; |
218 return result; |
516 } |
219 } |
517 |
220 |
518 // == Macro Expansion == |
221 bool ZBarrierSetC2::array_copy_requires_gc_barriers(bool tightly_coupled_alloc, BasicType type, |
519 |
222 bool is_clone, ArrayCopyPhase phase) const { |
520 // Optimized, low spill, loadbarrier variant using stub specialized on register used |
223 return type == T_OBJECT || type == T_ARRAY; |
521 void ZBarrierSetC2::expand_loadbarrier_node(PhaseMacroExpand* phase, LoadBarrierNode* barrier) const { |
224 } |
522 PhaseIterGVN &igvn = phase->igvn(); |
225 |
523 float unlikely = PROB_UNLIKELY(0.999); |
226 // == Dominating barrier elision == |
524 |
227 |
525 Node* in_ctrl = barrier->in(LoadBarrierNode::Control); |
228 static bool block_has_safepoint(const Block* block, uint from, uint to) { |
526 Node* in_mem = barrier->in(LoadBarrierNode::Memory); |
229 for (uint i = from; i < to; i++) { |
527 Node* in_val = barrier->in(LoadBarrierNode::Oop); |
230 if (block->get_node(i)->is_MachSafePoint()) { |
528 Node* in_adr = barrier->in(LoadBarrierNode::Address); |
231 // Safepoint found |
529 |
232 return true; |
530 Node* out_ctrl = barrier->proj_out(LoadBarrierNode::Control); |
233 } |
531 Node* out_res = barrier->proj_out(LoadBarrierNode::Oop); |
234 } |
532 |
235 |
533 assert(barrier->in(LoadBarrierNode::Oop) != NULL, "oop to loadbarrier node cannot be null"); |
236 // Safepoint not found |
534 |
237 return false; |
535 Node* jthread = igvn.transform(new ThreadLocalNode()); |
238 } |
536 Node* adr = phase->basic_plus_adr(jthread, in_bytes(ZThreadLocalData::address_bad_mask_offset())); |
239 |
537 Node* bad_mask = igvn.transform(LoadNode::make(igvn, in_ctrl, in_mem, adr, |
240 static bool block_has_safepoint(const Block* block) { |
538 TypeRawPtr::BOTTOM, TypeX_X, TypeX_X->basic_type(), |
241 return block_has_safepoint(block, 0, block->number_of_nodes()); |
539 MemNode::unordered)); |
242 } |
540 Node* cast = igvn.transform(new CastP2XNode(in_ctrl, in_val)); |
243 |
541 Node* obj_masked = igvn.transform(new AndXNode(cast, bad_mask)); |
244 static uint block_index(const Block* block, const Node* node) { |
542 Node* cmp = igvn.transform(new CmpXNode(obj_masked, igvn.zerocon(TypeX_X->basic_type()))); |
245 for (uint j = 0; j < block->number_of_nodes(); ++j) { |
543 Node *bol = igvn.transform(new BoolNode(cmp, BoolTest::ne))->as_Bool(); |
246 if (block->get_node(j) == node) { |
544 IfNode* iff = igvn.transform(new IfNode(in_ctrl, bol, unlikely, COUNT_UNKNOWN))->as_If(); |
247 return j; |
545 Node* then = igvn.transform(new IfTrueNode(iff)); |
248 } |
546 Node* elsen = igvn.transform(new IfFalseNode(iff)); |
249 } |
547 |
250 ShouldNotReachHere(); |
548 Node* new_loadp = igvn.transform(new LoadBarrierSlowRegNode(then, in_adr, in_val, |
251 return 0; |
549 (const TypePtr*) in_val->bottom_type(), barrier->is_weak())); |
252 } |
550 |
253 |
551 // Create the final region/phi pair to converge cntl/data paths to downstream code |
254 void ZBarrierSetC2::analyze_dominating_barriers() const { |
552 Node* result_region = igvn.transform(new RegionNode(3)); |
255 ResourceMark rm; |
553 result_region->set_req(1, then); |
256 Compile* const C = Compile::current(); |
554 result_region->set_req(2, elsen); |
257 PhaseCFG* const cfg = C->cfg(); |
555 |
258 Block_List worklist; |
556 Node* result_phi = igvn.transform(new PhiNode(result_region, TypeInstPtr::BOTTOM)); |
259 Node_List mem_ops; |
557 result_phi->set_req(1, new_loadp); |
260 Node_List barrier_loads; |
558 result_phi->set_req(2, barrier->in(LoadBarrierNode::Oop)); |
261 |
559 |
262 // Step 1 - Find accesses, and track them in lists |
560 igvn.replace_node(out_ctrl, result_region); |
263 for (uint i = 0; i < cfg->number_of_blocks(); ++i) { |
561 igvn.replace_node(out_res, result_phi); |
264 const Block* const block = cfg->get_block(i); |
562 |
265 for (uint j = 0; j < block->number_of_nodes(); ++j) { |
563 assert(barrier->outcnt() == 0,"LoadBarrier macro node has non-null outputs after expansion!"); |
266 const Node* const node = block->get_node(j); |
564 |
267 if (!node->is_Mach()) { |
565 igvn.remove_dead_node(barrier); |
|
566 igvn.remove_dead_node(out_ctrl); |
|
567 igvn.remove_dead_node(out_res); |
|
568 |
|
569 assert(is_gc_barrier_node(result_phi), "sanity"); |
|
570 assert(step_over_gc_barrier(result_phi) == in_val, "sanity"); |
|
571 |
|
572 phase->C->print_method(PHASE_BARRIER_EXPANSION, 4, barrier->_idx); |
|
573 } |
|
574 |
|
575 bool ZBarrierSetC2::expand_barriers(Compile* C, PhaseIterGVN& igvn) const { |
|
576 ZBarrierSetC2State* s = state(); |
|
577 if (s->load_barrier_count() > 0) { |
|
578 PhaseMacroExpand macro(igvn); |
|
579 |
|
580 int skipped = 0; |
|
581 while (s->load_barrier_count() > skipped) { |
|
582 int load_barrier_count = s->load_barrier_count(); |
|
583 LoadBarrierNode * n = s->load_barrier_node(load_barrier_count-1-skipped); |
|
584 if (igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) { |
|
585 // Node is unreachable, so don't try to expand it |
|
586 s->remove_load_barrier_node(n); |
|
587 continue; |
268 continue; |
588 } |
269 } |
589 if (!n->can_be_eliminated()) { |
270 |
590 skipped++; |
271 MachNode* const mach = node->as_Mach(); |
|
272 switch (mach->ideal_Opcode()) { |
|
273 case Op_LoadP: |
|
274 case Op_CompareAndExchangeP: |
|
275 case Op_CompareAndSwapP: |
|
276 case Op_GetAndSetP: |
|
277 if (mach->barrier_data() == ZLoadBarrierStrong) { |
|
278 barrier_loads.push(mach); |
|
279 } |
|
280 case Op_StoreP: |
|
281 mem_ops.push(mach); |
|
282 break; |
|
283 |
|
284 default: |
|
285 break; |
|
286 } |
|
287 } |
|
288 } |
|
289 |
|
290 // Step 2 - Find dominating accesses for each load |
|
291 for (uint i = 0; i < barrier_loads.size(); i++) { |
|
292 MachNode* const load = barrier_loads.at(i)->as_Mach(); |
|
293 const TypePtr* load_adr_type = NULL; |
|
294 intptr_t load_offset = 0; |
|
295 const Node* const load_obj = load->get_base_and_disp(load_offset, load_adr_type); |
|
296 Block* const load_block = cfg->get_block_for_node(load); |
|
297 const uint load_index = block_index(load_block, load); |
|
298 |
|
299 for (uint j = 0; j < mem_ops.size(); j++) { |
|
300 MachNode* mem = mem_ops.at(j)->as_Mach(); |
|
301 const TypePtr* mem_adr_type = NULL; |
|
302 intptr_t mem_offset = 0; |
|
303 const Node* mem_obj = mem_obj = mem->get_base_and_disp(mem_offset, mem_adr_type); |
|
304 Block* mem_block = cfg->get_block_for_node(mem); |
|
305 uint mem_index = block_index(mem_block, mem); |
|
306 |
|
307 if (load_obj == NodeSentinel || mem_obj == NodeSentinel || |
|
308 load_obj == NULL || mem_obj == NULL || |
|
309 load_offset < 0 || mem_offset < 0) { |
591 continue; |
310 continue; |
592 } |
311 } |
593 expand_loadbarrier_node(¯o, n); |
312 |
594 assert(s->load_barrier_count() < load_barrier_count, "must have deleted a node from load barrier list"); |
313 if (mem_obj != load_obj || mem_offset != load_offset) { |
595 if (C->failing()) { |
314 // Not the same addresses, not a candidate |
596 return true; |
315 continue; |
597 } |
316 } |
598 } |
317 |
599 while (s->load_barrier_count() > 0) { |
318 if (load_block == mem_block) { |
600 int load_barrier_count = s->load_barrier_count(); |
319 // Earlier accesses in the same block |
601 LoadBarrierNode* n = s->load_barrier_node(load_barrier_count - 1); |
320 if (mem_index < load_index && !block_has_safepoint(mem_block, mem_index + 1, load_index)) { |
602 assert(!(igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())), "should have been processed already"); |
321 load->set_barrier_data(ZLoadBarrierElided); |
603 assert(!n->can_be_eliminated(), "should have been processed already"); |
322 } |
604 expand_loadbarrier_node(¯o, n); |
323 } else if (mem_block->dominates(load_block)) { |
605 assert(s->load_barrier_count() < load_barrier_count, "must have deleted a node from load barrier list"); |
324 // Dominating block? Look around for safepoints |
606 if (C->failing()) { |
325 ResourceMark rm; |
607 return true; |
326 Block_List stack; |
608 } |
327 VectorSet visited(Thread::current()->resource_area()); |
609 } |
328 stack.push(load_block); |
610 igvn.set_delay_transform(false); |
329 bool safepoint_found = block_has_safepoint(load_block); |
611 igvn.optimize(); |
330 while (!safepoint_found && stack.size() > 0) { |
612 if (C->failing()) { |
331 Block* block = stack.pop(); |
613 return true; |
332 if (visited.test_set(block->_pre_order)) { |
614 } |
333 continue; |
615 } |
334 } |
616 |
335 if (block_has_safepoint(block)) { |
617 return false; |
336 safepoint_found = true; |
618 } |
337 break; |
619 |
338 } |
620 Node* ZBarrierSetC2::step_over_gc_barrier(Node* c) const { |
339 if (block == mem_block) { |
621 Node* node = c; |
340 continue; |
622 |
341 } |
623 // 1. This step follows potential oop projections of a load barrier before expansion |
342 |
624 if (node->is_Proj()) { |
343 // Push predecessor blocks |
625 node = node->in(0); |
344 for (uint p = 1; p < block->num_preds(); ++p) { |
626 } |
345 Block* pred = cfg->get_block_for_node(block->pred(p)); |
627 |
346 stack.push(pred); |
628 // 2. This step checks for unexpanded load barriers |
347 } |
629 if (node->is_LoadBarrier()) { |
348 } |
630 return node->in(LoadBarrierNode::Oop); |
349 |
631 } |
350 if (!safepoint_found) { |
632 |
351 load->set_barrier_data(ZLoadBarrierElided); |
633 // 3. This step checks for the phi corresponding to an optimized load barrier expansion |
352 } |
634 if (node->is_Phi()) { |
353 } |
635 PhiNode* phi = node->as_Phi(); |
354 } |
636 Node* n = phi->in(1); |
355 } |
637 if (n != NULL && n->is_LoadBarrierSlowReg()) { |
356 } |
638 assert(c == node, "projections from step 1 should only be seen before macro expansion"); |
357 |
639 return phi->in(2); |
358 // == Reduced spilling optimization == |
640 } |
359 |
641 } |
360 void ZBarrierSetC2::compute_liveness_at_stubs() const { |
642 |
|
643 return c; |
|
644 } |
|
645 |
|
646 Node* ZBarrierSetC2::step_over_gc_barrier_ctrl(Node* c) const { |
|
647 Node* node = c; |
|
648 |
|
649 // 1. This step follows potential ctrl projections of a load barrier before expansion |
|
650 if (node->is_Proj()) { |
|
651 node = node->in(0); |
|
652 } |
|
653 |
|
654 // 2. This step checks for unexpanded load barriers |
|
655 if (node->is_LoadBarrier()) { |
|
656 return node->in(LoadBarrierNode::Control); |
|
657 } |
|
658 |
|
659 return c; |
|
660 } |
|
661 |
|
662 bool ZBarrierSetC2::array_copy_requires_gc_barriers(bool tightly_coupled_alloc, BasicType type, bool is_clone, ArrayCopyPhase phase) const { |
|
663 return is_reference_type(type); |
|
664 } |
|
665 |
|
666 bool ZBarrierSetC2::final_graph_reshaping(Compile* compile, Node* n, uint opcode) const { |
|
667 switch (opcode) { |
|
668 case Op_LoadBarrier: |
|
669 assert(0, "There should be no load barriers left"); |
|
670 case Op_ZGetAndSetP: |
|
671 case Op_ZCompareAndExchangeP: |
|
672 case Op_ZCompareAndSwapP: |
|
673 case Op_ZWeakCompareAndSwapP: |
|
674 #ifdef ASSERT |
|
675 if (VerifyOptoOopOffsets) { |
|
676 MemNode *mem = n->as_Mem(); |
|
677 // Check to see if address types have grounded out somehow. |
|
678 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); |
|
679 ciInstanceKlass *k = tp->klass()->as_instance_klass(); |
|
680 bool oop_offset_is_sane = k->contains_field_offset(tp->offset()); |
|
681 assert(!tp || oop_offset_is_sane, ""); |
|
682 } |
|
683 #endif |
|
684 return true; |
|
685 default: |
|
686 return false; |
|
687 } |
|
688 } |
|
689 |
|
690 bool ZBarrierSetC2::matcher_find_shared_visit(Matcher* matcher, Matcher::MStack& mstack, Node* n, uint opcode, bool& mem_op, int& mem_addr_idx) const { |
|
691 switch(opcode) { |
|
692 case Op_CallLeaf: |
|
693 if (n->as_Call()->entry_point() == ZBarrierSetRuntime::load_barrier_on_oop_field_preloaded_addr() || |
|
694 n->as_Call()->entry_point() == ZBarrierSetRuntime::load_barrier_on_weak_oop_field_preloaded_addr()) { |
|
695 mem_op = true; |
|
696 mem_addr_idx = TypeFunc::Parms + 1; |
|
697 return true; |
|
698 } |
|
699 return false; |
|
700 default: |
|
701 return false; |
|
702 } |
|
703 } |
|
704 |
|
705 bool ZBarrierSetC2::matcher_find_shared_post_visit(Matcher* matcher, Node* n, uint opcode) const { |
|
706 switch(opcode) { |
|
707 case Op_ZCompareAndExchangeP: |
|
708 case Op_ZCompareAndSwapP: |
|
709 case Op_ZWeakCompareAndSwapP: { |
|
710 Node *mem = n->in(MemNode::Address); |
|
711 Node *keepalive = n->in(5); |
|
712 Node *pair1 = new BinaryNode(mem, keepalive); |
|
713 |
|
714 Node *newval = n->in(MemNode::ValueIn); |
|
715 Node *oldval = n->in(LoadStoreConditionalNode::ExpectedIn); |
|
716 Node *pair2 = new BinaryNode(oldval, newval); |
|
717 |
|
718 n->set_req(MemNode::Address, pair1); |
|
719 n->set_req(MemNode::ValueIn, pair2); |
|
720 n->del_req(5); |
|
721 n->del_req(LoadStoreConditionalNode::ExpectedIn); |
|
722 return true; |
|
723 } |
|
724 case Op_ZGetAndSetP: { |
|
725 Node *keepalive = n->in(4); |
|
726 Node *newval = n->in(MemNode::ValueIn); |
|
727 Node *pair = new BinaryNode(newval, keepalive); |
|
728 n->set_req(MemNode::ValueIn, pair); |
|
729 n->del_req(4); |
|
730 return true; |
|
731 } |
|
732 |
|
733 default: |
|
734 return false; |
|
735 } |
|
736 } |
|
737 |
|
738 // == Verification == |
|
739 |
|
740 #ifdef ASSERT |
|
741 |
|
742 static void verify_slippery_safepoints_internal(Node* ctrl) { |
|
743 // Given a CFG node, make sure it does not contain both safepoints and loads |
|
744 // that have expanded barriers. |
|
745 bool found_safepoint = false; |
|
746 bool found_load = false; |
|
747 |
|
748 for (DUIterator_Fast imax, i = ctrl->fast_outs(imax); i < imax; i++) { |
|
749 Node* node = ctrl->fast_out(i); |
|
750 if (node->in(0) != ctrl) { |
|
751 // Skip outgoing precedence edges from ctrl. |
|
752 continue; |
|
753 } |
|
754 if (node->is_SafePoint()) { |
|
755 found_safepoint = true; |
|
756 } |
|
757 if (node->is_Load() && load_require_barrier(node->as_Load()) && |
|
758 load_has_expanded_barrier(node->as_Load())) { |
|
759 found_load = true; |
|
760 } |
|
761 } |
|
762 assert(!found_safepoint || !found_load, "found load and safepoint in same block"); |
|
763 } |
|
764 |
|
765 static void verify_slippery_safepoints(Compile* C) { |
|
766 ResourceArea *area = Thread::current()->resource_area(); |
|
767 Unique_Node_List visited(area); |
|
768 Unique_Node_List checked(area); |
|
769 |
|
770 // Recursively walk the graph. |
|
771 visited.push(C->root()); |
|
772 while (visited.size() > 0) { |
|
773 Node* node = visited.pop(); |
|
774 |
|
775 Node* ctrl = node; |
|
776 if (!node->is_CFG()) { |
|
777 ctrl = node->in(0); |
|
778 } |
|
779 |
|
780 if (ctrl != NULL && !checked.member(ctrl)) { |
|
781 // For each block found in the graph, verify that it does not |
|
782 // contain both a safepoint and a load requiring barriers. |
|
783 verify_slippery_safepoints_internal(ctrl); |
|
784 |
|
785 checked.push(ctrl); |
|
786 } |
|
787 |
|
788 checked.push(node); |
|
789 |
|
790 for (DUIterator_Fast imax, i = node->fast_outs(imax); i < imax; i++) { |
|
791 Node* use = node->fast_out(i); |
|
792 if (checked.member(use)) continue; |
|
793 if (visited.member(use)) continue; |
|
794 visited.push(use); |
|
795 } |
|
796 } |
|
797 } |
|
798 |
|
799 void ZBarrierSetC2::verify_gc_barriers(Compile* compile, CompilePhase phase) const { |
|
800 switch(phase) { |
|
801 case BarrierSetC2::BeforeOptimize: |
|
802 case BarrierSetC2::BeforeLateInsertion: |
|
803 assert(state()->load_barrier_count() == 0, "No barriers inserted yet"); |
|
804 break; |
|
805 case BarrierSetC2::BeforeMacroExpand: |
|
806 // Barrier placement should be set by now. |
|
807 verify_gc_barriers(false /*post_parse*/); |
|
808 break; |
|
809 case BarrierSetC2::BeforeCodeGen: |
|
810 // Barriers has been fully expanded. |
|
811 assert(state()->load_barrier_count() == 0, "No more macro barriers"); |
|
812 verify_slippery_safepoints(compile); |
|
813 break; |
|
814 default: |
|
815 assert(0, "Phase without verification"); |
|
816 } |
|
817 } |
|
818 |
|
819 // post_parse implies that there might be load barriers without uses after parsing |
|
820 // That only applies when adding barriers at parse time. |
|
821 void ZBarrierSetC2::verify_gc_barriers(bool post_parse) const { |
|
822 ZBarrierSetC2State* s = state(); |
|
823 Compile* C = Compile::current(); |
|
824 ResourceMark rm; |
361 ResourceMark rm; |
825 VectorSet visited(Thread::current()->resource_area()); |
362 Compile* const C = Compile::current(); |
826 |
363 Arena* const A = Thread::current()->resource_area(); |
827 for (int i = 0; i < s->load_barrier_count(); i++) { |
364 PhaseCFG* const cfg = C->cfg(); |
828 LoadBarrierNode* n = s->load_barrier_node(i); |
365 PhaseRegAlloc* const regalloc = C->regalloc(); |
829 |
366 RegMask* const live = NEW_ARENA_ARRAY(A, RegMask, cfg->number_of_blocks() * sizeof(RegMask)); |
830 // The dominating barrier on the same address if it exists and |
367 ZBarrierSetAssembler* const bs = ZBarrierSet::assembler(); |
831 // this barrier must not be applied on the value from the same |
368 Block_List worklist; |
832 // load otherwise the value is not reloaded before it's used the |
369 |
833 // second time. |
370 for (uint i = 0; i < cfg->number_of_blocks(); ++i) { |
834 assert(n->in(LoadBarrierNode::Similar)->is_top() || |
371 new ((void*)(live + i)) RegMask(); |
835 (n->in(LoadBarrierNode::Similar)->in(0)->is_LoadBarrier() && |
372 worklist.push(cfg->get_block(i)); |
836 n->in(LoadBarrierNode::Similar)->in(0)->in(LoadBarrierNode::Address) == n->in(LoadBarrierNode::Address) && |
373 } |
837 n->in(LoadBarrierNode::Similar)->in(0)->in(LoadBarrierNode::Oop) != n->in(LoadBarrierNode::Oop)), |
374 |
838 "broken similar edge"); |
375 while (worklist.size() > 0) { |
839 |
376 const Block* const block = worklist.pop(); |
840 assert(n->as_LoadBarrier()->has_true_uses(), |
377 RegMask& old_live = live[block->_pre_order]; |
841 "found unneeded load barrier"); |
378 RegMask new_live; |
842 |
379 |
843 // Several load barrier nodes chained through their Similar edge |
380 // Initialize to union of successors |
844 // break the code that remove the barriers in final graph reshape. |
381 for (uint i = 0; i < block->_num_succs; i++) { |
845 assert(n->in(LoadBarrierNode::Similar)->is_top() || |
382 const uint succ_id = block->_succs[i]->_pre_order; |
846 (n->in(LoadBarrierNode::Similar)->in(0)->is_LoadBarrier() && |
383 new_live.OR(live[succ_id]); |
847 n->in(LoadBarrierNode::Similar)->in(0)->in(LoadBarrierNode::Similar)->is_top()), |
384 } |
848 "chain of Similar load barriers"); |
385 |
849 |
386 // Walk block backwards, computing liveness |
850 if (!n->in(LoadBarrierNode::Similar)->is_top()) { |
387 for (int i = block->number_of_nodes() - 1; i >= 0; --i) { |
851 ResourceMark rm; |
388 const Node* const node = block->get_node(i); |
852 Unique_Node_List wq; |
389 |
853 Node* other = n->in(LoadBarrierNode::Similar)->in(0); |
390 // Remove def bits |
854 wq.push(n); |
391 const OptoReg::Name first = bs->refine_register(node, regalloc->get_reg_first(node)); |
855 for (uint next = 0; next < wq.size(); ++next) { |
392 const OptoReg::Name second = bs->refine_register(node, regalloc->get_reg_second(node)); |
856 Node *nn = wq.at(next); |
393 if (first != OptoReg::Bad) { |
857 assert(nn->is_CFG(), ""); |
394 new_live.Remove(first); |
858 assert(!nn->is_SafePoint(), ""); |
395 } |
859 |
396 if (second != OptoReg::Bad) { |
860 if (nn == other) { |
397 new_live.Remove(second); |
861 continue; |
398 } |
862 } |
399 |
863 |
400 // Add use bits |
864 if (nn->is_Region()) { |
401 for (uint j = 1; j < node->req(); ++j) { |
865 for (uint i = 1; i < nn->req(); i++) { |
402 const Node* const use = node->in(j); |
866 Node* m = nn->in(i); |
403 const OptoReg::Name first = bs->refine_register(use, regalloc->get_reg_first(use)); |
867 if (m != NULL) { |
404 const OptoReg::Name second = bs->refine_register(use, regalloc->get_reg_second(use)); |
868 wq.push(m); |
405 if (first != OptoReg::Bad) { |
869 } |
406 new_live.Insert(first); |
870 } |
407 } |
871 } else { |
408 if (second != OptoReg::Bad) { |
872 Node* m = nn->in(0); |
409 new_live.Insert(second); |
873 if (m != NULL) { |
410 } |
874 wq.push(m); |
411 } |
875 } |
412 |
876 } |
413 // If this node tracks liveness, update it |
877 } |
414 RegMask* const regs = barrier_set_state()->live(node); |
878 } |
415 if (regs != NULL) { |
879 } |
416 regs->OR(new_live); |
880 } |
417 } |
881 |
418 } |
882 #endif // end verification code |
419 |
883 |
420 // Now at block top, see if we have any changes |
884 // If a call is the control, we actually want its control projection |
421 new_live.SUBTRACT(old_live); |
885 static Node* normalize_ctrl(Node* node) { |
422 if (new_live.is_NotEmpty()) { |
886 if (node->is_Call()) { |
423 // Liveness has refined, update and propagate to prior blocks |
887 node = node->as_Call()->proj_out(TypeFunc::Control); |
424 old_live.OR(new_live); |
888 } |
425 for (uint i = 1; i < block->num_preds(); ++i) { |
889 return node; |
426 Block* const pred = cfg->get_block_for_node(block->pred(i)); |
890 } |
427 worklist.push(pred); |
891 |
428 } |
892 static Node* get_ctrl_normalized(PhaseIdealLoop *phase, Node* node) { |
429 } |
893 return normalize_ctrl(phase->get_ctrl(node)); |
430 } |
894 } |
431 } |
895 |
|
896 static void call_catch_cleanup_one(PhaseIdealLoop* phase, LoadNode* load, Node* ctrl); |
|
897 |
|
898 // This code is cloning all uses of a load that is between a call and the catch blocks, |
|
899 // to each use. |
|
900 |
|
901 static bool fixup_uses_in_catch(PhaseIdealLoop *phase, Node *start_ctrl, Node *node) { |
|
902 |
|
903 if (!phase->has_ctrl(node)) { |
|
904 // This node is floating - doesn't need to be cloned. |
|
905 assert(node != start_ctrl, "check"); |
|
906 return false; |
|
907 } |
|
908 |
|
909 Node* ctrl = get_ctrl_normalized(phase, node); |
|
910 if (ctrl != start_ctrl) { |
|
911 // We are in a successor block - the node is ok. |
|
912 return false; // Unwind |
|
913 } |
|
914 |
|
915 // Process successor nodes |
|
916 int outcnt = node->outcnt(); |
|
917 for (int i = 0; i < outcnt; i++) { |
|
918 Node* n = node->raw_out(0); |
|
919 assert(!n->is_LoadBarrier(), "Sanity"); |
|
920 // Calling recursively, visiting leafs first |
|
921 fixup_uses_in_catch(phase, start_ctrl, n); |
|
922 } |
|
923 |
|
924 // Now all successors are outside |
|
925 // - Clone this node to both successors |
|
926 assert(!node->is_Store(), "Stores not expected here"); |
|
927 |
|
928 // In some very rare cases a load that doesn't need a barrier will end up here |
|
929 // Treat it as a LoadP and the insertion of phis will be done correctly. |
|
930 if (node->is_Load()) { |
|
931 call_catch_cleanup_one(phase, node->as_Load(), phase->get_ctrl(node)); |
|
932 } else { |
|
933 for (DUIterator_Fast jmax, i = node->fast_outs(jmax); i < jmax; i++) { |
|
934 Node* use = node->fast_out(i); |
|
935 Node* clone = node->clone(); |
|
936 assert(clone->outcnt() == 0, ""); |
|
937 |
|
938 assert(use->find_edge(node) != -1, "check"); |
|
939 phase->igvn().rehash_node_delayed(use); |
|
940 use->replace_edge(node, clone); |
|
941 |
|
942 Node* new_ctrl; |
|
943 if (use->is_block_start()) { |
|
944 new_ctrl = use; |
|
945 } else if (use->is_CFG()) { |
|
946 new_ctrl = use->in(0); |
|
947 assert (new_ctrl != NULL, ""); |
|
948 } else { |
|
949 new_ctrl = get_ctrl_normalized(phase, use); |
|
950 } |
|
951 |
|
952 phase->set_ctrl(clone, new_ctrl); |
|
953 |
|
954 if (phase->C->directive()->ZTraceLoadBarriersOption) tty->print_cr(" Clone op %i as %i to control %i", node->_idx, clone->_idx, new_ctrl->_idx); |
|
955 phase->igvn().register_new_node_with_optimizer(clone); |
|
956 --i, --jmax; |
|
957 } |
|
958 assert(node->outcnt() == 0, "must be empty now"); |
|
959 |
|
960 // Node node is dead. |
|
961 phase->igvn().remove_dead_node(node); |
|
962 } |
|
963 return true; // unwind - return if a use was processed |
|
964 } |
|
965 |
|
966 // Clone a load to a specific catch_proj |
|
967 static Node* clone_load_to_catchproj(PhaseIdealLoop* phase, Node* load, Node* catch_proj) { |
|
968 Node* cloned_load = load->clone(); |
|
969 cloned_load->set_req(0, catch_proj); // set explicit control |
|
970 phase->set_ctrl(cloned_load, catch_proj); // update |
|
971 if (phase->C->directive()->ZTraceLoadBarriersOption) tty->print_cr(" Clone LOAD %i as %i to control %i", load->_idx, cloned_load->_idx, catch_proj->_idx); |
|
972 phase->igvn().register_new_node_with_optimizer(cloned_load); |
|
973 return cloned_load; |
|
974 } |
|
975 |
|
976 static Node* get_dominating_region(PhaseIdealLoop* phase, Node* node, Node* stop) { |
|
977 Node* region = node; |
|
978 while (!region->isa_Region()) { |
|
979 Node *up = phase->idom(region); |
|
980 assert(up != region, "Must not loop"); |
|
981 assert(up != stop, "Must not find original control"); |
|
982 region = up; |
|
983 } |
|
984 return region; |
|
985 } |
|
986 |
|
987 // Clone this load to each catch block |
|
988 static void call_catch_cleanup_one(PhaseIdealLoop* phase, LoadNode* load, Node* ctrl) { |
|
989 bool trace = phase->C->directive()->ZTraceLoadBarriersOption; |
|
990 phase->igvn().set_delay_transform(true); |
|
991 |
|
992 // Verify pre conditions |
|
993 assert(ctrl->isa_Proj() && ctrl->in(0)->isa_Call(), "Must be a call proj"); |
|
994 assert(ctrl->raw_out(0)->isa_Catch(), "Must be a catch"); |
|
995 |
|
996 if (ctrl->raw_out(0)->isa_Catch()->outcnt() == 1) { |
|
997 if (trace) tty->print_cr("Cleaning up catch: Skipping load %i, call with single catch", load->_idx); |
|
998 return; |
|
999 } |
|
1000 |
|
1001 // Process the loads successor nodes - if any is between |
|
1002 // the call and the catch blocks, they need to be cloned to. |
|
1003 // This is done recursively |
|
1004 for (uint i = 0; i < load->outcnt();) { |
|
1005 Node *n = load->raw_out(i); |
|
1006 assert(!n->is_LoadBarrier(), "Sanity"); |
|
1007 if (!fixup_uses_in_catch(phase, ctrl, n)) { |
|
1008 // if no successor was cloned, progress to next out. |
|
1009 i++; |
|
1010 } |
|
1011 } |
|
1012 |
|
1013 // Now all the loads uses has been cloned down |
|
1014 // Only thing left is to clone the loads, but they must end up |
|
1015 // first in the catch blocks. |
|
1016 |
|
1017 // We clone the loads oo the catch blocks only when needed. |
|
1018 // An array is used to map the catch blocks to each lazily cloned load. |
|
1019 // In that way no extra unnecessary loads are cloned. |
|
1020 |
|
1021 // Any use dominated by original block must have an phi and a region added |
|
1022 |
|
1023 Node* catch_node = ctrl->raw_out(0); |
|
1024 int number_of_catch_projs = catch_node->outcnt(); |
|
1025 Node** proj_to_load_mapping = NEW_RESOURCE_ARRAY(Node*, number_of_catch_projs); |
|
1026 Copy::zero_to_bytes(proj_to_load_mapping, sizeof(Node*) * number_of_catch_projs); |
|
1027 |
|
1028 // The phi_map is used to keep track of where phis have already been inserted |
|
1029 int phi_map_len = phase->C->unique(); |
|
1030 Node** phi_map = NEW_RESOURCE_ARRAY(Node*, phi_map_len); |
|
1031 Copy::zero_to_bytes(phi_map, sizeof(Node*) * phi_map_len); |
|
1032 |
|
1033 for (unsigned int i = 0; i < load->outcnt(); i++) { |
|
1034 Node* load_use_control = NULL; |
|
1035 Node* load_use = load->raw_out(i); |
|
1036 |
|
1037 if (phase->has_ctrl(load_use)) { |
|
1038 load_use_control = get_ctrl_normalized(phase, load_use); |
|
1039 assert(load_use_control != ctrl, "sanity"); |
|
1040 } else { |
|
1041 load_use_control = load_use->in(0); |
|
1042 } |
|
1043 assert(load_use_control != NULL, "sanity"); |
|
1044 if (trace) tty->print_cr(" Handling use: %i, with control: %i", load_use->_idx, load_use_control->_idx); |
|
1045 |
|
1046 // Some times the loads use is a phi. For them we need to determine from which catch block |
|
1047 // the use is defined. |
|
1048 bool load_use_is_phi = false; |
|
1049 unsigned int load_use_phi_index = 0; |
|
1050 Node* phi_ctrl = NULL; |
|
1051 if (load_use->is_Phi()) { |
|
1052 // Find phi input that matches load |
|
1053 for (unsigned int u = 1; u < load_use->req(); u++) { |
|
1054 if (load_use->in(u) == load) { |
|
1055 load_use_is_phi = true; |
|
1056 load_use_phi_index = u; |
|
1057 assert(load_use->in(0)->is_Region(), "Region or broken"); |
|
1058 phi_ctrl = load_use->in(0)->in(u); |
|
1059 assert(phi_ctrl->is_CFG(), "check"); |
|
1060 assert(phi_ctrl != load, "check"); |
|
1061 break; |
|
1062 } |
|
1063 } |
|
1064 assert(load_use_is_phi, "must find"); |
|
1065 assert(load_use_phi_index > 0, "sanity"); |
|
1066 } |
|
1067 |
|
1068 // For each load use, see which catch projs dominates, create load clone lazily and reconnect |
|
1069 bool found_dominating_catchproj = false; |
|
1070 for (int c = 0; c < number_of_catch_projs; c++) { |
|
1071 Node* catchproj = catch_node->raw_out(c); |
|
1072 assert(catchproj != NULL && catchproj->isa_CatchProj(), "Sanity"); |
|
1073 |
|
1074 if (!phase->is_dominator(catchproj, load_use_control)) { |
|
1075 if (load_use_is_phi && phase->is_dominator(catchproj, phi_ctrl)) { |
|
1076 // The loads use is local to the catchproj. |
|
1077 // fall out and replace load with catch-local load clone. |
|
1078 } else { |
|
1079 continue; |
|
1080 } |
|
1081 } |
|
1082 assert(!found_dominating_catchproj, "Max one should match"); |
|
1083 |
|
1084 // Clone loads to catch projs |
|
1085 Node* load_clone = proj_to_load_mapping[c]; |
|
1086 if (load_clone == NULL) { |
|
1087 load_clone = clone_load_to_catchproj(phase, load, catchproj); |
|
1088 proj_to_load_mapping[c] = load_clone; |
|
1089 } |
|
1090 phase->igvn().rehash_node_delayed(load_use); |
|
1091 |
|
1092 if (load_use_is_phi) { |
|
1093 // phis are special - the load is defined from a specific control flow |
|
1094 load_use->set_req(load_use_phi_index, load_clone); |
|
1095 } else { |
|
1096 // Multipe edges can be replaced at once - on calls for example |
|
1097 load_use->replace_edge(load, load_clone); |
|
1098 } |
|
1099 --i; // more than one edge can have been removed, but the next is in later iterations |
|
1100 |
|
1101 // We could break the for-loop after finding a dominating match. |
|
1102 // But keep iterating to catch any bad idom early. |
|
1103 found_dominating_catchproj = true; |
|
1104 } |
|
1105 |
|
1106 // We found no single catchproj that dominated the use - The use is at a point after |
|
1107 // where control flow from multiple catch projs have merged. We will have to create |
|
1108 // phi nodes before the use and tie the output from the cloned loads together. It |
|
1109 // can be a single phi or a number of chained phis, depending on control flow |
|
1110 if (!found_dominating_catchproj) { |
|
1111 |
|
1112 // Use phi-control if use is a phi |
|
1113 if (load_use_is_phi) { |
|
1114 load_use_control = phi_ctrl; |
|
1115 } |
|
1116 assert(phase->is_dominator(ctrl, load_use_control), "Common use but no dominator"); |
|
1117 |
|
1118 // Clone a load on all paths |
|
1119 for (int c = 0; c < number_of_catch_projs; c++) { |
|
1120 Node* catchproj = catch_node->raw_out(c); |
|
1121 Node* load_clone = proj_to_load_mapping[c]; |
|
1122 if (load_clone == NULL) { |
|
1123 load_clone = clone_load_to_catchproj(phase, load, catchproj); |
|
1124 proj_to_load_mapping[c] = load_clone; |
|
1125 } |
|
1126 } |
|
1127 |
|
1128 // Move up dominator tree from use until dom front is reached |
|
1129 Node* next_region = get_dominating_region(phase, load_use_control, ctrl); |
|
1130 while (phase->idom(next_region) != catch_node) { |
|
1131 next_region = phase->idom(next_region); |
|
1132 if (trace) tty->print_cr("Moving up idom to region ctrl %i", next_region->_idx); |
|
1133 } |
|
1134 assert(phase->is_dominator(catch_node, next_region), "Sanity"); |
|
1135 |
|
1136 // Create or reuse phi node that collect all cloned loads and feed it to the use. |
|
1137 Node* test_phi = phi_map[next_region->_idx]; |
|
1138 if ((test_phi != NULL) && test_phi->is_Phi()) { |
|
1139 // Reuse an already created phi |
|
1140 if (trace) tty->print_cr(" Using cached Phi %i on load_use %i", test_phi->_idx, load_use->_idx); |
|
1141 phase->igvn().rehash_node_delayed(load_use); |
|
1142 load_use->replace_edge(load, test_phi); |
|
1143 // Now this use is done |
|
1144 } else { |
|
1145 // Otherwise we need to create one or more phis |
|
1146 PhiNode* next_phi = new PhiNode(next_region, load->type()); |
|
1147 phi_map[next_region->_idx] = next_phi; // cache new phi |
|
1148 phase->igvn().rehash_node_delayed(load_use); |
|
1149 load_use->replace_edge(load, next_phi); |
|
1150 |
|
1151 int dominators_of_region = 0; |
|
1152 do { |
|
1153 // New phi, connect to region and add all loads as in. |
|
1154 Node* region = next_region; |
|
1155 assert(region->isa_Region() && region->req() > 2, "Catch dead region nodes"); |
|
1156 PhiNode* new_phi = next_phi; |
|
1157 |
|
1158 if (trace) tty->print_cr("Created Phi %i on load %i with control %i", new_phi->_idx, load->_idx, region->_idx); |
|
1159 |
|
1160 // Need to add all cloned loads to the phi, taking care that the right path is matched |
|
1161 dominators_of_region = 0; // reset for new region |
|
1162 for (unsigned int reg_i = 1; reg_i < region->req(); reg_i++) { |
|
1163 Node* region_pred = region->in(reg_i); |
|
1164 assert(region_pred->is_CFG(), "check"); |
|
1165 bool pred_has_dominator = false; |
|
1166 for (int c = 0; c < number_of_catch_projs; c++) { |
|
1167 Node* catchproj = catch_node->raw_out(c); |
|
1168 if (phase->is_dominator(catchproj, region_pred)) { |
|
1169 new_phi->set_req(reg_i, proj_to_load_mapping[c]); |
|
1170 if (trace) tty->print_cr(" - Phi in(%i) set to load %i", reg_i, proj_to_load_mapping[c]->_idx); |
|
1171 pred_has_dominator = true; |
|
1172 dominators_of_region++; |
|
1173 break; |
|
1174 } |
|
1175 } |
|
1176 |
|
1177 // Sometimes we need to chain several phis. |
|
1178 if (!pred_has_dominator) { |
|
1179 assert(dominators_of_region <= 1, "More than one region can't require extra phi"); |
|
1180 if (trace) tty->print_cr(" - Region %i pred %i not dominated by catch proj", region->_idx, region_pred->_idx); |
|
1181 // Continue search on on this region_pred |
|
1182 // - walk up to next region |
|
1183 // - create a new phi and connect to first new_phi |
|
1184 next_region = get_dominating_region(phase, region_pred, ctrl); |
|
1185 |
|
1186 // Lookup if there already is a phi, create a new otherwise |
|
1187 Node* test_phi = phi_map[next_region->_idx]; |
|
1188 if ((test_phi != NULL) && test_phi->is_Phi()) { |
|
1189 next_phi = test_phi->isa_Phi(); |
|
1190 dominators_of_region++; // record that a match was found and that we are done |
|
1191 if (trace) tty->print_cr(" Using cached phi Phi %i on control %i", next_phi->_idx, next_region->_idx); |
|
1192 } else { |
|
1193 next_phi = new PhiNode(next_region, load->type()); |
|
1194 phi_map[next_region->_idx] = next_phi; |
|
1195 } |
|
1196 new_phi->set_req(reg_i, next_phi); |
|
1197 } |
|
1198 } |
|
1199 |
|
1200 new_phi->set_req(0, region); |
|
1201 phase->igvn().register_new_node_with_optimizer(new_phi); |
|
1202 phase->set_ctrl(new_phi, region); |
|
1203 |
|
1204 assert(dominators_of_region != 0, "Must have found one this iteration"); |
|
1205 } while (dominators_of_region == 1); |
|
1206 } |
|
1207 --i; |
|
1208 } |
|
1209 } // end of loop over uses |
|
1210 |
|
1211 assert(load->outcnt() == 0, "All uses should be handled"); |
|
1212 phase->igvn().remove_dead_node(load); |
|
1213 phase->C->print_method(PHASE_CALL_CATCH_CLEANUP, 4, load->_idx); |
|
1214 |
|
1215 // Now we should be home |
|
1216 phase->igvn().set_delay_transform(false); |
|
1217 } |
|
1218 |
|
1219 // Sort out the loads that are between a call ant its catch blocks |
|
1220 static void process_catch_cleanup_candidate(PhaseIdealLoop* phase, LoadNode* load, bool verify) { |
|
1221 bool trace = phase->C->directive()->ZTraceLoadBarriersOption; |
|
1222 |
|
1223 Node* ctrl = get_ctrl_normalized(phase, load); |
|
1224 if (!ctrl->is_Proj() || (ctrl->in(0) == NULL) || !ctrl->in(0)->isa_Call()) { |
|
1225 return; |
|
1226 } |
|
1227 |
|
1228 Node* catch_node = ctrl->isa_Proj()->raw_out(0); |
|
1229 if (catch_node->is_Catch()) { |
|
1230 if (catch_node->outcnt() > 1) { |
|
1231 assert(!verify, "All loads should already have been moved"); |
|
1232 call_catch_cleanup_one(phase, load, ctrl); |
|
1233 } else { |
|
1234 if (trace) tty->print_cr("Call catch cleanup with only one catch: load %i ", load->_idx); |
|
1235 } |
|
1236 } |
|
1237 } |
|
1238 |
|
1239 void ZBarrierSetC2::barrier_insertion_phase(Compile* C, PhaseIterGVN& igvn) const { |
|
1240 PhaseIdealLoop::optimize(igvn, LoopOptsZBarrierInsertion); |
|
1241 if (C->failing()) return; |
|
1242 } |
|
1243 |
|
1244 bool ZBarrierSetC2::optimize_loops(PhaseIdealLoop* phase, LoopOptsMode mode, VectorSet& visited, Node_Stack& nstack, Node_List& worklist) const { |
|
1245 |
|
1246 if (mode == LoopOptsZBarrierInsertion) { |
|
1247 // First make sure all loads between call and catch are moved to the catch block |
|
1248 clean_catch_blocks(phase); |
|
1249 DEBUG_ONLY(clean_catch_blocks(phase, true /* verify */);) |
|
1250 |
|
1251 // Then expand barriers on all loads |
|
1252 insert_load_barriers(phase); |
|
1253 |
|
1254 // Handle all Unsafe that need barriers. |
|
1255 insert_barriers_on_unsafe(phase); |
|
1256 |
|
1257 phase->C->clear_major_progress(); |
|
1258 return true; |
|
1259 } else { |
|
1260 return false; |
|
1261 } |
|
1262 } |
|
1263 |
|
1264 static bool can_simplify_cas(LoadStoreNode* node) { |
|
1265 if (node->isa_LoadStoreConditional()) { |
|
1266 Node *expected_in = node->as_LoadStoreConditional()->in(LoadStoreConditionalNode::ExpectedIn); |
|
1267 return (expected_in->get_ptr_type() == TypePtr::NULL_PTR); |
|
1268 } else { |
|
1269 return false; |
|
1270 } |
|
1271 } |
|
1272 |
|
1273 static void insert_barrier_before_unsafe(PhaseIdealLoop* phase, LoadStoreNode* old_node) { |
|
1274 |
|
1275 Compile *C = phase->C; |
|
1276 PhaseIterGVN &igvn = phase->igvn(); |
|
1277 LoadStoreNode* zclone = NULL; |
|
1278 |
|
1279 Node *in_ctrl = old_node->in(MemNode::Control); |
|
1280 Node *in_mem = old_node->in(MemNode::Memory); |
|
1281 Node *in_adr = old_node->in(MemNode::Address); |
|
1282 Node *in_val = old_node->in(MemNode::ValueIn); |
|
1283 const TypePtr *adr_type = old_node->adr_type(); |
|
1284 const TypePtr* load_type = TypeOopPtr::BOTTOM; // The type for the load we are adding |
|
1285 |
|
1286 switch (old_node->Opcode()) { |
|
1287 case Op_CompareAndExchangeP: { |
|
1288 zclone = new ZCompareAndExchangePNode(in_ctrl, in_mem, in_adr, in_val, old_node->in(LoadStoreConditionalNode::ExpectedIn), |
|
1289 adr_type, old_node->get_ptr_type(), ((CompareAndExchangeNode*)old_node)->order()); |
|
1290 load_type = old_node->bottom_type()->is_ptr(); |
|
1291 break; |
|
1292 } |
|
1293 case Op_WeakCompareAndSwapP: { |
|
1294 if (can_simplify_cas(old_node)) { |
|
1295 break; |
|
1296 } |
|
1297 zclone = new ZWeakCompareAndSwapPNode(in_ctrl, in_mem, in_adr, in_val, old_node->in(LoadStoreConditionalNode::ExpectedIn), |
|
1298 ((CompareAndSwapNode*)old_node)->order()); |
|
1299 adr_type = TypePtr::BOTTOM; |
|
1300 break; |
|
1301 } |
|
1302 case Op_CompareAndSwapP: { |
|
1303 if (can_simplify_cas(old_node)) { |
|
1304 break; |
|
1305 } |
|
1306 zclone = new ZCompareAndSwapPNode(in_ctrl, in_mem, in_adr, in_val, old_node->in(LoadStoreConditionalNode::ExpectedIn), |
|
1307 ((CompareAndSwapNode*)old_node)->order()); |
|
1308 adr_type = TypePtr::BOTTOM; |
|
1309 break; |
|
1310 } |
|
1311 case Op_GetAndSetP: { |
|
1312 zclone = new ZGetAndSetPNode(in_ctrl, in_mem, in_adr, in_val, old_node->adr_type(), old_node->get_ptr_type()); |
|
1313 load_type = old_node->bottom_type()->is_ptr(); |
|
1314 break; |
|
1315 } |
|
1316 } |
|
1317 if (zclone != NULL) { |
|
1318 igvn.register_new_node_with_optimizer(zclone, old_node); |
|
1319 |
|
1320 // Make load |
|
1321 LoadPNode *load = new LoadPNode(NULL, in_mem, in_adr, adr_type, load_type, MemNode::unordered, |
|
1322 LoadNode::DependsOnlyOnTest); |
|
1323 load_set_expanded_barrier(load); |
|
1324 igvn.register_new_node_with_optimizer(load); |
|
1325 igvn.replace_node(old_node, zclone); |
|
1326 |
|
1327 Node *barrier = new LoadBarrierNode(C, NULL, in_mem, load, in_adr, false /* weak */); |
|
1328 Node *barrier_val = new ProjNode(barrier, LoadBarrierNode::Oop); |
|
1329 Node *barrier_ctrl = new ProjNode(barrier, LoadBarrierNode::Control); |
|
1330 |
|
1331 igvn.register_new_node_with_optimizer(barrier); |
|
1332 igvn.register_new_node_with_optimizer(barrier_val); |
|
1333 igvn.register_new_node_with_optimizer(barrier_ctrl); |
|
1334 |
|
1335 // loop over all of in_ctrl usages and move to barrier_ctrl |
|
1336 for (DUIterator_Last imin, i = in_ctrl->last_outs(imin); i >= imin; --i) { |
|
1337 Node *use = in_ctrl->last_out(i); |
|
1338 uint l; |
|
1339 for (l = 0; use->in(l) != in_ctrl; l++) {} |
|
1340 igvn.replace_input_of(use, l, barrier_ctrl); |
|
1341 } |
|
1342 |
|
1343 load->set_req(MemNode::Control, in_ctrl); |
|
1344 barrier->set_req(LoadBarrierNode::Control, in_ctrl); |
|
1345 zclone->add_req(barrier_val); // add req as keep alive. |
|
1346 |
|
1347 C->print_method(PHASE_ADD_UNSAFE_BARRIER, 4, zclone->_idx); |
|
1348 } |
|
1349 } |
|
1350 |
|
1351 void ZBarrierSetC2::insert_barriers_on_unsafe(PhaseIdealLoop* phase) const { |
|
1352 Compile *C = phase->C; |
|
1353 PhaseIterGVN &igvn = phase->igvn(); |
|
1354 uint new_ids = C->unique(); |
|
1355 VectorSet visited(Thread::current()->resource_area()); |
|
1356 GrowableArray<Node *> nodeStack(Thread::current()->resource_area(), 0, 0, NULL); |
|
1357 nodeStack.push(C->root()); |
|
1358 visited.test_set(C->root()->_idx); |
|
1359 |
|
1360 // Traverse all nodes, visit all unsafe ops that require a barrier |
|
1361 while (nodeStack.length() > 0) { |
|
1362 Node *n = nodeStack.pop(); |
|
1363 |
|
1364 bool is_old_node = (n->_idx < new_ids); // don't process nodes that were created during cleanup |
|
1365 if (is_old_node) { |
|
1366 if (n->is_LoadStore()) { |
|
1367 LoadStoreNode* lsn = n->as_LoadStore(); |
|
1368 if (lsn->has_barrier()) { |
|
1369 BasicType bt = lsn->in(MemNode::Address)->bottom_type()->basic_type(); |
|
1370 assert (is_reference_type(bt), "Sanity test"); |
|
1371 insert_barrier_before_unsafe(phase, lsn); |
|
1372 } |
|
1373 } |
|
1374 } |
|
1375 for (uint i = 0; i < n->len(); i++) { |
|
1376 if (n->in(i)) { |
|
1377 if (!visited.test_set(n->in(i)->_idx)) { |
|
1378 nodeStack.push(n->in(i)); |
|
1379 } |
|
1380 } |
|
1381 } |
|
1382 } |
|
1383 |
|
1384 igvn.optimize(); |
|
1385 C->print_method(PHASE_ADD_UNSAFE_BARRIER, 2); |
|
1386 } |
|
1387 |
|
1388 // The purpose of ZBarrierSetC2::clean_catch_blocks is to prepare the IR for |
|
1389 // splicing in load barrier nodes. |
|
1390 // |
|
1391 // The problem is that we might have instructions between a call and its catch nodes. |
|
1392 // (This is usually handled in PhaseCFG:call_catch_cleanup, which clones mach nodes in |
|
1393 // already scheduled blocks.) We can't have loads that require barriers there, |
|
1394 // because we need to splice in new control flow, and that would violate the IR. |
|
1395 // |
|
1396 // clean_catch_blocks find all Loads that require a barrier and clone them and any |
|
1397 // dependent instructions to each use. The loads must be in the beginning of the catch block |
|
1398 // before any store. |
|
1399 // |
|
1400 // Sometimes the loads use will be at a place dominated by all catch blocks, then we need |
|
1401 // a load in each catch block, and a Phi at the dominated use. |
|
1402 |
|
1403 void ZBarrierSetC2::clean_catch_blocks(PhaseIdealLoop* phase, bool verify) const { |
|
1404 |
|
1405 Compile *C = phase->C; |
|
1406 uint new_ids = C->unique(); |
|
1407 PhaseIterGVN &igvn = phase->igvn(); |
|
1408 VectorSet visited(Thread::current()->resource_area()); |
|
1409 GrowableArray<Node *> nodeStack(Thread::current()->resource_area(), 0, 0, NULL); |
|
1410 nodeStack.push(C->root()); |
|
1411 visited.test_set(C->root()->_idx); |
|
1412 |
|
1413 // Traverse all nodes, visit all loads that require a barrier |
|
1414 while(nodeStack.length() > 0) { |
|
1415 Node *n = nodeStack.pop(); |
|
1416 |
|
1417 for (uint i = 0; i < n->len(); i++) { |
|
1418 if (n->in(i)) { |
|
1419 if (!visited.test_set(n->in(i)->_idx)) { |
|
1420 nodeStack.push(n->in(i)); |
|
1421 } |
|
1422 } |
|
1423 } |
|
1424 |
|
1425 bool is_old_node = (n->_idx < new_ids); // don't process nodes that were created during cleanup |
|
1426 if (n->is_Load() && is_old_node) { |
|
1427 LoadNode* load = n->isa_Load(); |
|
1428 // only care about loads that will have a barrier |
|
1429 if (load_require_barrier(load)) { |
|
1430 process_catch_cleanup_candidate(phase, load, verify); |
|
1431 } |
|
1432 } |
|
1433 } |
|
1434 |
|
1435 C->print_method(PHASE_CALL_CATCH_CLEANUP, 2); |
|
1436 } |
|
1437 |
|
1438 class DomDepthCompareClosure : public CompareClosure<LoadNode*> { |
|
1439 PhaseIdealLoop* _phase; |
|
1440 |
|
1441 public: |
|
1442 DomDepthCompareClosure(PhaseIdealLoop* phase) : _phase(phase) { } |
|
1443 |
|
1444 int do_compare(LoadNode* const &n1, LoadNode* const &n2) { |
|
1445 int d1 = _phase->dom_depth(_phase->get_ctrl(n1)); |
|
1446 int d2 = _phase->dom_depth(_phase->get_ctrl(n2)); |
|
1447 if (d1 == d2) { |
|
1448 // Compare index if the depth is the same, ensures all entries are unique. |
|
1449 return n1->_idx - n2->_idx; |
|
1450 } else { |
|
1451 return d2 - d1; |
|
1452 } |
|
1453 } |
|
1454 }; |
|
1455 |
|
1456 // Traverse graph and add all loadPs to list, sorted by dom depth |
|
1457 void gather_loadnodes_sorted(PhaseIdealLoop* phase, GrowableArray<LoadNode*>* loadList) { |
|
1458 |
|
1459 VectorSet visited(Thread::current()->resource_area()); |
|
1460 GrowableArray<Node *> nodeStack(Thread::current()->resource_area(), 0, 0, NULL); |
|
1461 DomDepthCompareClosure ddcc(phase); |
|
1462 |
|
1463 nodeStack.push(phase->C->root()); |
|
1464 while(nodeStack.length() > 0) { |
|
1465 Node *n = nodeStack.pop(); |
|
1466 if (visited.test(n->_idx)) { |
|
1467 continue; |
|
1468 } |
|
1469 |
|
1470 if (n->isa_Load()) { |
|
1471 LoadNode *load = n->as_Load(); |
|
1472 if (load_require_barrier(load)) { |
|
1473 assert(phase->get_ctrl(load) != NULL, "sanity"); |
|
1474 assert(phase->dom_depth(phase->get_ctrl(load)) != 0, "sanity"); |
|
1475 loadList->insert_sorted(&ddcc, load); |
|
1476 } |
|
1477 } |
|
1478 |
|
1479 visited.set(n->_idx); |
|
1480 for (uint i = 0; i < n->req(); i++) { |
|
1481 if (n->in(i)) { |
|
1482 if (!visited.test(n->in(i)->_idx)) { |
|
1483 nodeStack.push(n->in(i)); |
|
1484 } |
|
1485 } |
|
1486 } |
|
1487 } |
|
1488 } |
|
1489 |
|
1490 // Add LoadBarriers to all LoadPs |
|
1491 void ZBarrierSetC2::insert_load_barriers(PhaseIdealLoop* phase) const { |
|
1492 |
|
1493 bool trace = phase->C->directive()->ZTraceLoadBarriersOption; |
|
1494 GrowableArray<LoadNode *> loadList(Thread::current()->resource_area(), 0, 0, NULL); |
|
1495 gather_loadnodes_sorted(phase, &loadList); |
|
1496 |
|
1497 PhaseIterGVN &igvn = phase->igvn(); |
|
1498 int count = 0; |
|
1499 |
|
1500 for (GrowableArrayIterator<LoadNode *> loadIter = loadList.begin(); loadIter != loadList.end(); ++loadIter) { |
|
1501 LoadNode *load = *loadIter; |
|
1502 |
|
1503 if (load_has_expanded_barrier(load)) { |
|
1504 continue; |
|
1505 } |
|
1506 |
|
1507 do { |
|
1508 // Insert a barrier on a loadP |
|
1509 // if another load is found that needs to be expanded first, retry on that one |
|
1510 LoadNode* result = insert_one_loadbarrier(phase, load, phase->get_ctrl(load)); |
|
1511 while (result != NULL) { |
|
1512 result = insert_one_loadbarrier(phase, result, phase->get_ctrl(result)); |
|
1513 } |
|
1514 } while (!load_has_expanded_barrier(load)); |
|
1515 } |
|
1516 |
|
1517 phase->C->print_method(PHASE_INSERT_BARRIER, 2); |
|
1518 } |
|
1519 |
|
1520 void push_antidependent_stores(PhaseIdealLoop* phase, Node_Stack& nodestack, LoadNode* start_load) { |
|
1521 // push all stores on the same mem, that can_alias |
|
1522 // Any load found must be handled first |
|
1523 PhaseIterGVN &igvn = phase->igvn(); |
|
1524 int load_alias_idx = igvn.C->get_alias_index(start_load->adr_type()); |
|
1525 |
|
1526 Node *mem = start_load->in(1); |
|
1527 for (DUIterator_Fast imax, u = mem->fast_outs(imax); u < imax; u++) { |
|
1528 Node *mem_use = mem->fast_out(u); |
|
1529 |
|
1530 if (mem_use == start_load) continue; |
|
1531 if (!mem_use->is_Store()) continue; |
|
1532 if (!phase->has_ctrl(mem_use)) continue; |
|
1533 if (phase->get_ctrl(mem_use) != phase->get_ctrl(start_load)) continue; |
|
1534 |
|
1535 // add any aliasing store in this block |
|
1536 StoreNode *store = mem_use->isa_Store(); |
|
1537 const TypePtr *adr_type = store->adr_type(); |
|
1538 if (igvn.C->can_alias(adr_type, load_alias_idx)) { |
|
1539 nodestack.push(store, 0); |
|
1540 } |
|
1541 } |
|
1542 } |
|
1543 |
|
1544 LoadNode* ZBarrierSetC2::insert_one_loadbarrier(PhaseIdealLoop* phase, LoadNode* start_load, Node* ctrl) const { |
|
1545 bool trace = phase->C->directive()->ZTraceLoadBarriersOption; |
|
1546 PhaseIterGVN &igvn = phase->igvn(); |
|
1547 |
|
1548 // Check for other loadPs at the same loop depth that is reachable by a DFS |
|
1549 // - if found - return it. It needs to be inserted first |
|
1550 // - otherwise proceed and insert barrier |
|
1551 |
|
1552 VectorSet visited(Thread::current()->resource_area()); |
|
1553 Node_Stack nodestack(100); |
|
1554 |
|
1555 nodestack.push(start_load, 0); |
|
1556 push_antidependent_stores(phase, nodestack, start_load); |
|
1557 |
|
1558 while(!nodestack.is_empty()) { |
|
1559 Node* n = nodestack.node(); // peek |
|
1560 nodestack.pop(); |
|
1561 if (visited.test(n->_idx)) { |
|
1562 continue; |
|
1563 } |
|
1564 |
|
1565 if (n->is_Load() && n != start_load && load_require_barrier(n->as_Load()) && !load_has_expanded_barrier(n->as_Load())) { |
|
1566 // Found another load that needs a barrier in the same block. Must expand later loads first. |
|
1567 if (trace) tty->print_cr(" * Found LoadP %i on DFS", n->_idx); |
|
1568 return n->as_Load(); // return node that should be expanded first |
|
1569 } |
|
1570 |
|
1571 if (!phase->has_ctrl(n)) continue; |
|
1572 if (phase->get_ctrl(n) != phase->get_ctrl(start_load)) continue; |
|
1573 if (n->is_Phi()) continue; |
|
1574 |
|
1575 visited.set(n->_idx); |
|
1576 // push all children |
|
1577 for (DUIterator_Fast imax, ii = n->fast_outs(imax); ii < imax; ii++) { |
|
1578 Node* c = n->fast_out(ii); |
|
1579 if (c != NULL) { |
|
1580 nodestack.push(c, 0); |
|
1581 } |
|
1582 } |
|
1583 } |
|
1584 |
|
1585 insert_one_loadbarrier_inner(phase, start_load, ctrl, visited); |
|
1586 return NULL; |
|
1587 } |
|
1588 |
|
1589 void ZBarrierSetC2::insert_one_loadbarrier_inner(PhaseIdealLoop* phase, LoadNode* load, Node* ctrl, VectorSet visited2) const { |
|
1590 PhaseIterGVN &igvn = phase->igvn(); |
|
1591 Compile* C = igvn.C; |
|
1592 bool trace = C->directive()->ZTraceLoadBarriersOption; |
|
1593 |
|
1594 // create barrier |
|
1595 Node* barrier = new LoadBarrierNode(C, NULL, load->in(LoadNode::Memory), NULL, load->in(LoadNode::Address), load_has_weak_barrier(load)); |
|
1596 Node* barrier_val = new ProjNode(barrier, LoadBarrierNode::Oop); |
|
1597 Node* barrier_ctrl = new ProjNode(barrier, LoadBarrierNode::Control); |
|
1598 ctrl = normalize_ctrl(ctrl); |
|
1599 |
|
1600 if (trace) tty->print_cr("Insert load %i with barrier: %i and ctrl : %i", load->_idx, barrier->_idx, ctrl->_idx); |
|
1601 |
|
1602 // Splice control |
|
1603 // - insert barrier control diamond between loads ctrl and ctrl successor on path to block end. |
|
1604 // - If control successor is a catch, step over to next. |
|
1605 Node* ctrl_succ = NULL; |
|
1606 for (DUIterator_Fast imax, j = ctrl->fast_outs(imax); j < imax; j++) { |
|
1607 Node* tmp = ctrl->fast_out(j); |
|
1608 |
|
1609 // - CFG nodes is the ones we are going to splice (1 only!) |
|
1610 // - Phi nodes will continue to hang from the region node! |
|
1611 // - self loops should be skipped |
|
1612 if (tmp->is_Phi() || tmp == ctrl) { |
|
1613 continue; |
|
1614 } |
|
1615 |
|
1616 if (tmp->is_CFG()) { |
|
1617 assert(ctrl_succ == NULL, "There can be only one"); |
|
1618 ctrl_succ = tmp; |
|
1619 continue; |
|
1620 } |
|
1621 } |
|
1622 |
|
1623 // Now splice control |
|
1624 assert(ctrl_succ != load, "sanity"); |
|
1625 assert(ctrl_succ != NULL, "Broken IR"); |
|
1626 bool found = false; |
|
1627 for(uint k = 0; k < ctrl_succ->req(); k++) { |
|
1628 if (ctrl_succ->in(k) == ctrl) { |
|
1629 assert(!found, "sanity"); |
|
1630 if (trace) tty->print_cr(" Move CFG ctrl_succ %i to barrier_ctrl", ctrl_succ->_idx); |
|
1631 igvn.replace_input_of(ctrl_succ, k, barrier_ctrl); |
|
1632 found = true; |
|
1633 k--; |
|
1634 } |
|
1635 } |
|
1636 |
|
1637 // For all successors of ctrl - move all visited to become successors of barrier_ctrl instead |
|
1638 for (DUIterator_Fast imax, r = ctrl->fast_outs(imax); r < imax; r++) { |
|
1639 Node* tmp = ctrl->fast_out(r); |
|
1640 if (tmp->is_SafePoint() || (visited2.test(tmp->_idx) && (tmp != load))) { |
|
1641 if (trace) tty->print_cr(" Move ctrl_succ %i to barrier_ctrl", tmp->_idx); |
|
1642 igvn.replace_input_of(tmp, 0, barrier_ctrl); |
|
1643 --r; --imax; |
|
1644 } |
|
1645 } |
|
1646 |
|
1647 // Move the loads user to the barrier |
|
1648 for (DUIterator_Fast imax, i = load->fast_outs(imax); i < imax; i++) { |
|
1649 Node* u = load->fast_out(i); |
|
1650 if (u->isa_LoadBarrier()) { |
|
1651 continue; |
|
1652 } |
|
1653 |
|
1654 // find correct input - replace with iterator? |
|
1655 for(uint j = 0; j < u->req(); j++) { |
|
1656 if (u->in(j) == load) { |
|
1657 igvn.replace_input_of(u, j, barrier_val); |
|
1658 --i; --imax; // Adjust the iterator of the *outer* loop |
|
1659 break; // some nodes (calls) might have several uses from the same node |
|
1660 } |
|
1661 } |
|
1662 } |
|
1663 |
|
1664 // Connect barrier to load and control |
|
1665 barrier->set_req(LoadBarrierNode::Oop, load); |
|
1666 barrier->set_req(LoadBarrierNode::Control, ctrl); |
|
1667 |
|
1668 igvn.replace_input_of(load, MemNode::Control, ctrl); |
|
1669 load->pin(); |
|
1670 |
|
1671 igvn.rehash_node_delayed(load); |
|
1672 igvn.register_new_node_with_optimizer(barrier); |
|
1673 igvn.register_new_node_with_optimizer(barrier_val); |
|
1674 igvn.register_new_node_with_optimizer(barrier_ctrl); |
|
1675 load_set_expanded_barrier(load); |
|
1676 |
|
1677 C->print_method(PHASE_INSERT_BARRIER, 3, load->_idx); |
|
1678 } |
|
1679 |
|
1680 // The bad_mask in the ThreadLocalData shouldn't have an anti-dep-check. |
|
1681 // The bad_mask address if of type TypeRawPtr, but that will alias |
|
1682 // InitializeNodes until the type system is expanded. |
|
1683 bool ZBarrierSetC2::needs_anti_dependence_check(const Node* node) const { |
|
1684 MachNode* mnode = node->as_Mach(); |
|
1685 if (mnode != NULL) { |
|
1686 intptr_t offset = 0; |
|
1687 const TypePtr *adr_type2 = NULL; |
|
1688 const Node* base = mnode->get_base_and_disp(offset, adr_type2); |
|
1689 if ((base != NULL) && |
|
1690 (base->is_Mach() && base->as_Mach()->ideal_Opcode() == Op_ThreadLocal) && |
|
1691 (offset == in_bytes(ZThreadLocalData::address_bad_mask_offset()))) { |
|
1692 return false; |
|
1693 } |
|
1694 } |
|
1695 return true; |
|
1696 } |
|